A Study of the Learnability of Relational Properties: Model Counting Meets Machine Learning (MCML)
This paper introduces the MCML approach for empirically studying
the learnability of relational properties that can be
expressed in the well-known software design language Alloy. A
key novelty of MCML is quantification of the performance of and semantic
differences among trained machine learning (ML) models, specifically
decision trees, with respect to entire (bounded) input spaces,
and not just for given training and test
datasets (as is the common practice). MCML reduces the quantification
problems to the classic complexity theory problem of model counting, and employs state-of-the-art model
counters. The results show that relatively simple
ML models can achieve surprisingly high performance (accuracy and F1-score)
when evaluated in the common setting of using
training and test datasets – even when the training dataset is
much smaller than the test dataset – indicating the seeming
simplicity of learning relational properties. However, MCML
metrics based on model counting show that the performance can degrade
substantially when tested against the entire (bounded) input space,
indicating the high complexity of precisely learning these properties,
and the usefulness of model counting in quantifying the true performance.
Thu 18 JunDisplayed time zone: Pacific Time (US & Canada) change
10:40 - 11:40 | Machine Learning IIPLDI Research Papers at PLDI Research Papers live stream Chair(s): Ke Wang Visa Research | ||
10:40 20mTalk | Proving Data-Poisoning Robustness in Decision Trees PLDI Research Papers Samuel Drews University of Wisconsin-Madison, USA, Aws Albarghouthi University of Wisconsin-Madison, USA, Loris D'Antoni University of Wisconsin-Madison, USA | ||
11:00 20mTalk | A Study of the Learnability of Relational Properties: Model Counting Meets Machine Learning (MCML) PLDI Research Papers Muhammad Usman University of Texas at Austin, USA, Wenxi Wang University of Texas at Austin, USA, Marko Vasic University of Texas at Austin, USA, Kaiyuan Wang Google, USA, Haris Vikalo University of Texas at Austin, USA, Sarfraz Khurshid University of Texas at Austin, USA | ||
11:20 20mTalk | Learning Fast and Precise Numerical Analysis PLDI Research Papers Jingxuan He ETH Zurich, Switzerland, Gagandeep Singh ETH Zurich, Switzerland, Markus Püschel ETH Zurich, Switzerland, Martin Vechev ETH Zurich, Switzerland |